ABSTRACT
Due to the high spread of the COVID-19 virus, several diagnosis support systems are being developed in order to detect the disease in a faster and accurate way. In this paper, a stacking method for Computed Tomography (CT) scans has been implemented for the pre-processing step. The method combines both slice normalization and lung segmentation in a single output image using RGB color channels, providing more information from the input slices to the CNN models. The binary classification step starts with a slice-level prediction, which applies fine-tuning to the whole model and dense layers are changed by a custom scheme to improve the performance. Then, a patient-level prediction is performed by fixing a threshold percentage of COVID positive slices that allows to make the final prediction, classifying patients as COVID or NORMAL. The accuracy and metrics obtained show the robustness of the presented method in comparison to using the normalised slices or the masks independently. Given the results obtained, the proposed method can accurately detect the COVID-19 disease and the fusion of information improves the results obtained. © 2021, Springer Nature Switzerland AG.
ABSTRACT
Due to the urgency of the COVID pandemic, it is necessary to develop new and quick methods to detect the infection and stop the spread of the disease. In this work we compare a simple Deep Learning (DL) model with an ensemble model in the task of COVID detection in X-Ray images. For the simple model, we have used only frontal DX X-Ray images while, for the ensemble model, we have used frontal DX and CR X-Ray images, as well as lateral DX and CR X-Ray images. In the ensemble model, the features of the four images are combined to make a final prediction and, since not every patient possess all types of images, the model is also robust against missing information, which is crucial in these types of models. Although the dataset used is very noisy, the presented system has shown the desired robustness and offers relevant results, showing that ensemble models can generalize better over the data, which leads to a higher accuracy. Finally, we share our conclusions and discuss future work where we want to try using a similar methodology. © 2021, Springer Nature Switzerland AG.